PaperlayerView exact revision on arXiv

Source-linked AI summary

Online PCB Defect Detector On A New PCB Defect Dataset

Sanli Tang, Fan He, Xiaolin Huang, Jie Yang

arXiv:1902.06197v1cs.CV

TL;DR

PCB defect detection methods can fail on unaccounted patterns and sensitive hyper-parameters, while limited annotated data constrains advanced detector training. The paper introduces a GPP-based detector operating on template–tested image pairs and establishes DeepPCB with 1,500 annotated pairs. Experiments report improved detection performance and efficiency, while the conclusion describes state-of-the-art performance with low computational time.

  • Problem

    Existing image-difference and image-processing methods can fail on unaccounted defect patterns and hyper-parameter sensitivity, while insufficient annotated data limits advanced detector training.

  • Method

    The paper combines a template–tested image-pair detector with group pyramid pooling, which merges grouped features at multiple resolutions for scale-specific PCB defect prediction.

  • Results

    The proposed model improves mean average precision by 1.0% to 9.3% across compared methods and achieves higher F-score than the non-GPP and SSD-FPN alternatives.

  • Takeaways & Limitations

    DeepPCB provides a public, precisely annotated benchmark, while the GPP architecture combines multi-resolution features with low computational time for PCB defect detection.

Abstract

from arXiv · show

Previous works for PCB defect detection based on image difference and image processing techniques have already achieved promising performance. However, they sometimes fall short because of the unaccounted defect patterns or over-sensitivity about some hyper-parameters. In this work, we design a deep model that accurately detects PCB defects from an input pair of a detect-free template and a defective tested image. A novel group pyramid pooling module is proposed to efficiently extract features of a large range of resolutions, which are merged by group to predict PCB defect of corresponding scales. To train the deep model, a dataset is established, namely DeepPCB, which contains 1,500 image pairs with annotations including positions of 6 common types of PCB defects. Experiment results validate the effectiveness and efficiency of the proposed model by achieving $98.6\%$ mAP @ 62 FPS on DeepPCB dataset. This dataset is now available at: https://github.com/tangsanli5201/DeepPCB.

1. INTRODUCTION

PCB defect detection methods can struggle with complex patterns, image-pair misalignment, and hyper-parameter sensitivity, while limited annotated data constrains advanced model training. The paper addresses these issues by introducing the DeepPCB dataset and a group pyramid pooling-based detector.

  • Motivation: Annotated data scarcity limits validation of PCB defect detectors and training of advanced neural-network models.The dataset is intended to support research on detecting and classifying defects from template–tested image pairs.
  • Motivation: Image-difference and logic-based methods can fail on unaccounted defect patterns, image distortion or offset, and sensitive morphological hyper-parameters.The cited examples include erosion and dilation kernel sizes.
  • Dataset contribution: DeepPCB contains 1,500 aligned template–tested image pairs with position and class annotations for six PCB defect types.The dataset is described as the first public PCB defect dataset and uses template matching for alignment.
  • Model contribution: Group pyramid pooling merges grouped features at various resolutions, with each group combining local and larger-range context for corresponding defect scales.The module targets the accuracy–efficiency dilemma in deep PCB defect detectors.
  • Contributions: The paper reports three contributions: a GPP-based detector, the first dataset with 1,500 precisely annotated aligned pairs, and experiments evaluating both.The experiments assess the proposed model and the usefulness of DeepPCB.

2. THE DEEPPCB DATASET

DeepPCB is built from aligned PCB image pairs with six annotated defect categories, using standardized preprocessing and bounding-box labels. Its evaluation follows object-detection metrics with an IoU threshold for correctness.

  • Dataset composition: DeepPCB contains 1,500 PCB image pairs covering six defect types, with 1,000 pairs for training and 500 for testing.Each pair contains a 640 x 640 defect-free template and defective tested image.
  • Image preparation: The source images are clipped into 640 x 640 sub-images and aligned using template matching to reduce translation and rotation offsets.The original template and tested images are around 16k x 16k pixels.
  • Image preparation: Template matching and thresholding are common preprocessing techniques for high-accuracy PCB defect localization and classification.Thresholding is selected to reduce illumination disturbance, while specific preprocessing can vary by algorithm.
  • Annotations: Each defect is annotated with an axis-aligned bounding box and class ID across open, short, mousebite, spur, pin hole, and spurious copper categories.Artificial defects are added to tested images, producing around 3 to 12 defects per 640 x 640 image.
  • Evaluation: Average precision rate and F-mean evaluate detection, with correctness requiring same-class box IoU greater than 0.33 against a ground-truth box.The evaluation follows benchmarks for object and scene-text detection datasets.

3. APPROACH

The proposed detector compares features from a defect-free template and a tested image, then uses grouped pyramid pooling to predict defects at different scales. Default-box matching, regression, and classification losses train localization and six-class defect recognition.

  • Network Structure: The model extracts translation- and rotation-invariant features from both input images before calculating their differences.
  • Network Structure: Group pyramid pooling obtains multiple resolutions and assigns overlapping pooling groups to predict defects at specific scales.For example, the first group combines 1x1, 2x2, and 4x4 pooling features for small bounding boxes.
  • Prediction: Each default-box location predicts one of six PCB defect classes or background and estimates centroid offsets plus width and height scaling ratios.Default boxes use aspect ratios 0.5, 1.0, and 2.0, with sizes 0.04, 0.08, and 0.16 of the input image.
  • Objective Function: Ground-truth boxes are matched first to the maximum-overlap default box and then to default boxes whose Jaccard overlap exceeds 0.5.
  • Objective Function: Box regression uses a smooth L1 objective on normalized offsets between matched default and ground-truth boxes.The offsets represent center coordinates and width-height scaling relationships.
  • Objective Function: Classification uses softmax loss while sampling background boxes to maintain an approximately 3:1 background-to-foreground ratio.The background class index is 0.

4. EXPERIMENTS

Experiments evaluate PCB defect detection methods on DeepPCB, including image-processing, one-stage, and two-stage models, and assess the proposed GPP module through ablations. The proposed approach improves detection performance while maintaining efficient inference.

  • Results on DeepPCB Dataset: Evaluation compares image-processing, one-stage, and two-stage PCB defect detection methods on the DeepPCB dataset.The experiments also include ablation comparisons for the GPP module.
  • Results on DeepPCB Dataset: Training uses Adam on a single Titan X GPU for approximately 0.5 day.The reported setup uses an initial learning rate of 10^-3, 500 epochs, and batch size 16.
  • Results on DeepPCB Dataset: Figure 4 compares comparative models by inference speed in FPS and detection performance using F-score.The figure summarizes the speed-performance trade-off on DeepPCB.
  • Results on DeepPCB Dataset: The proposed model improves mAP by 1.0% to 9.3% over the compared methods on DeepPCB.The comparison includes one-stage models, a two-stage model, and an image-processing method.
  • Results on DeepPCB Dataset: 1.5% mAP improvement comes from using max pooling instead of average pooling in the GPP module.The pooling variants are identified as AP and MP in Table 1.
  • Ablation Study on Group Pyramid Pooling Module: The proposed model surpasses the non-GPP comparison by 3.9% F-score and SSD-FPN by 3.3% F-score.All compared models use the same VGG-tiny convolutional backbone.

5. CONCLUSION

The work contributes the DeepPCB dataset and the group pyramid pooling module for PCB defect detection. Experiments support the architecture’s state-of-the-art performance with low computational time, while the public dataset is intended to facilitate future research.

  • 5. CONCLUSION: DeepPCB provides PCB defect annotations for positions and six common defect types.The paper describes it as a large-scale dataset and the first dataset in terms of scale and precise annotations.
  • 5. CONCLUSION: Group pyramid pooling combines features at different resolutions for detecting PCB defects at various scales.The module is presented as an efficient deep feature-combination design.
  • 5. CONCLUSION: The proposed architecture achieves state-of-the-art performance while consuming very low computational time.This conclusion is based on extensive experiments.
  • 5. CONCLUSION: The public DeepPCB dataset is intended to facilitate future PCB defect detection research.The dataset is made available for research use.
Loading 1902.06197v1…